Multi-modal Intention Prediction with Probabilistic Movement Primitives

Abstract

This paper proposes a method for multi-modal prediction of intention based on a probabilistic description of movement primitives and goals. We target dyadic interaction between a human and a robot in a collaborative scenario. The robot acquires multi-modal models of collaborative action primitives containing gaze cues from the human partner and kinetic information about the manipulation primitives of its arm. We show that if the partner guides the robot with the gaze cue, the robot recognizes the intended action primitive even in the case of ambiguous actions. Furthermore, this prior knowledge acquired by gaze greatly improves the prediction of the future intended trajectory during a physical interaction. Results with the humanoid iCub are presented and discussed.